Energy-Aware Scheduling Algorithm Optimization for AI Workloads in Data Centers Based on Renewable Energy Supply Prediction

Abstract

This paper presents an innovative energy-aware scheduling algorithm that optimizes artificial intelligence workload distribution in data centers through advanced renewable energy supply forecasting. The proposed system integrates a hybrid LSTM-GRU neural network architecture, achieving a correlation coefficient of 0.87 for 24-hour renewable energy forecasts with a mean absolute error of 13%. Our priority-aware scheduling mechanism dynamically categorizes AI workloads based on energy intensity and deadline constraints, enabling optimal alignment with fluctuating renewable energy resources. Experimental evaluation across three geographically distributed data centers over a 12-month period demonstrates measurable improvements: 58% reduction in grid energy dependency, 47% decrease in carbon emissions, and 34% reduction in operational costs while maintaining 96.2% service level agreement compliance. The system architecture employs multi-objective optimization techniques, balancing energy efficiency, performance metrics, and carbon footprint considerations.